Higher-Order Occurrence Pooling for Bags-of-Words: Visual Concept Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• The Bag-of-Words (BoW) model is foundational in object recognition, relying on local descriptors and first-order occurrence pooling.
• Existing BoW models face limitations in capturing complex feature interactions and descriptor fusion.

Purpose of the Study:

• To investigate and develop higher-order pooling methods for the Bag-of-Words model.
• To enhance object recognition by effectively fusing diverse local image descriptors.

Main Methods:

• Derived Bag-of-Words with Higher-order Occurrence Pooling using linearized Minor Polynomial Kernel.
• Introduced higher-order pooling directly on local descriptors and a novel correlation-reducing pooling operator.
• Extended the model to accommodate various pooling operators and descriptor types.

Main Results:

• Demonstrated improved performance of First-, Second-, and Third-order Occurrence Pooling across various coders and pooling operators.
• Achieved superior results compared to established methods like Fisher Vector Encoding on standard benchmarks.
• Showcased effective fusion of multiple descriptor types using the proposed higher-order pooling framework.

Conclusions:

• Higher-order Occurrence Pooling offers a significant advancement over traditional first-order pooling in object recognition.
• The proposed methods provide a flexible and powerful framework for descriptor fusion and image representation.
• This work advances the state-of-the-art in object recognition by enhancing the Bag-of-Words model's representational capabilities.